Advanced background removal methods in single molecule localization microscopy using scattering networks and SVD

Accurate background estimation is a key challenge in single-molecule localization microscopy (SMLM), as it directly affects the quality of molecular localization and sample reconstruction. A coarse separation between background and relevant signal can often be obtained by contrasting spatial or temporal characteristics of the raw input. In this paper, we propose and compare two refined methods aimed at separating two types of background: (1) those where the variation over time occurs at a slower rate than the signal, and (2) backgrounds with distinctive spatial features. Filters that take advantage of Singular Value Decomposition (SVD) effectively address the first type of background, while the second can be managed using frequency-based filters. We introduce a novel approach based on neural networks to enhance background removal. Comparative evaluations using the Jaccard index (JI) demonstrate that the two methods improve localization performance, showing effective mitigation of background artifacts such as false positives or negatives, merged PSFs, and spurious localizations in SMLM data.